hi,bad result
when I use kijai wanwrapper node,the result is a red screen,nothing on it
when I use comfyui default ksampler,it's bad result,
@t8star I haven't tested this with kijai nodes.
Could you describe the bad results you're getting from the default ksampler?
yes,He flickers and blurs, and I asked other friends to test it as well. I have tried many samplers and schedulers, can you provide a correct workflow
@t8star I've just published a new LoRA that should fix this. It is a known limitation of the checkpoint because it was only fine-tuned on an image dataset so far. Therefore the videos it generates typically lack motion and quality degrades beyond the first frame.
The true solution is to add a large video dataset and continue fine-tuning, but curating that dataset will take a while and training will be resource intensive.
So, in the meantime, I trained a LoRA on a relatively small amount of videos to give this checkpoint some broad motion and temporal understanding of NSFW. Applying the LoRA should dramatically improve your results.
https://huggingface.co/NSFW-API/NSFW_Wan_1.3b_motion_helper
Please test with it and report back how it goes!
Thank you for your work. After using this LORA, the image is clearer and the content in JSON can be expressed. However, the human body is usually distorted, and so are the limbs. How else do you need to operate it? What are the recommended samplers, schedulers, and CFG?
I haven't tested with much other than a few basics.
CFG: 6
Shift: 8
Steps: 20-50
Scheduler: unipc
Resolution: 480p-720p
Sampler: WanVideoSampler in ComfyUI
You can test it out by loading it straight into the official Wan script and run inference through that, or through Musubi Tuner.
it often bad result,sometimes it provide a good result,not stability
@t8star You might be getting bad results from the prompt. This model was trained on data captioned using natural language rather than any sort of tags. Also, the model has no knowledge of the subreddit names themselves, so trigger words like "18_19" will be ineffective.
For a prompt like yours, I would try something like "Close-up, slightly low angle of a beautiful 19-year-old East Asian woman with long, dark hair, soaking in a bath. Her light skin glows against the white porcelain tub. She has large, perky breasts with small, dark areolas, and her shaved pussy and visible labia are just below the water's surface. She's looking directly at the viewer with a relaxed, sensual expression. The water glistens on her skin."
What works well for me is feeding the prompting guide into an LLM to use for reference and then describing the output I want and having it craft a good prompt.
would converting the new self-forcing 1.3 model similar to this one actually be better for quality in the base model?
good job
i train nsfw model
like me!
How to use it?